Grasp Planning from Human Prehension

TL;DR

This paper proposes a hierarchical grasp planner for robots, inspired by human hand movements. It breaks down grasp planning into two stages: first, mapping object and task information into a grasp-oriented description, and second, translating this into an "opposition space" that defines hand forces and degrees of freedom. The key innovation lies in using this opposition space and explicitly considering constraints to achieve more human-like grasping.

What Problem Does This Paper Address?

Designing effective grasp planners for robots is a significant challenge. While existing approaches often rely on hierarchical planning, this paper highlights the need to bridge the gap between high-level task specifications and the low-level details of how a robot hand should physically interact with an object. The authors draw inspiration from human prehension, observing that human grasping involves a complex interplay of object properties, task goals, and the hand's own capabilities. They aim to capture both the invariances (e.g., typical arm trajectories) and the variability (e.g., multiple grasping strategies) seen in human grasping to create more intelligent and adaptable robot grasp planners. A core problem is how to represent the object and task in a way that directly informs the hand's configuration and force application, and how to account for the numerous constraints that arise during this interaction.

Methods

The paper presents a two-level hierarchical grasp planner.

  1. Grasp-Oriented Task Requirements Generation: This first phase takes an object description (from an object database) and a task requirement (e.g., "swing up/down" for a hammer) and maps them into a more detailed set of "internal task requirements." These requirements encompass both functional constraints (how the object will be used) and physical constraints (forces, torques, etc., acting on the object).
  2. Opposition Space Generation: The second phase maps these internal task requirements into an "opposition space." This space is a novel representation that captures the available forces and degrees of freedom of the hand for a particular grasp. It's described in terms of "opposition vectors," which represent the axis along which force is applied between opposing surfaces (e.g., finger pads, palm). Virtual fingers (groups of real fingers) are used to define these oppositions. The resulting opposition space defines the functional capabilities of a specific hand posture for stable grasping and manipulation.

The authors emphasize that these mappings are not unique, and therefore, a crucial part of their approach is to identify and incorporate constraints acting on the hand/object/task interaction.

Datasets

Not discussed in the paper. The paper focuses on a conceptual framework and uses examples of objects (like a hammer) and their properties for illustration, rather than relying on specific datasets for training or evaluation.

Key Insights and Ideas

  • Human Prehension as a Model: The paper strongly advocates for studying human grasping to inform robot grasp planning, recognizing the sophistication and adaptability of the human hand.
  • Opposition Space: The introduction of "opposition space" as a representation for hand capabilities is a key contribution. It moves beyond simple joint angles to describe the functional potential of a grasp configuration in terms of forces and degrees of freedom.
  • Constraints are Crucial: The authors repeatedly stress that grasp planning is heavily constrained. These constraints arise from the hand's anatomy and biomechanics, the object's properties (size, shape, texture), and the dynamics of the task itself. Understanding and incorporating these constraints is vital for successful grasping.
  • Hierarchical Decomposition: The two-phase approach (task requirements -> opposition space) provides a structured way to break down the complex problem of grasp planning.

Key Results

The paper does not present quantitative experimental results. Instead, it offers a conceptual framework and a detailed analysis of the types of constraints that need to be considered. The "results" are the proposed planner architecture and the detailed breakdown of constraints, illustrated with examples like the hammer.

Limitations and Future Work

  • Lack of Empirical Verification: The authors explicitly state that empirical verification and quantification of the effects of constraints are "future tasks." The proposed planner is a theoretical framework.
  • Ignoring Higher-Level Constraints: The analysis focuses primarily on hand-object interaction. Constraints related to the arm and body, the laws of physics, and the central nervous system (CNS) are acknowledged as important but are "ignored" in this short analysis.
  • Mapping Uniqueness: The paper acknowledges that the mappings between object/task and grasp requirements, and between task requirements and opposition space, are not unique. Future work would involve developing methods to resolve these ambiguities using constraint knowledge.
  • Quantifying Constraints: A significant amount of further analysis is needed to quantify the various constraints identified.

Summary updated

BibTeX
@inproceedings{1987grasp,
  title = {{Grasp Planning from Human Prehension}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1987},
}
Figure